NemoClaw v0.0.64 is out!

[cv]

NemoClaw v0.0.64 is a stability release for the paths where sandboxes need to survive real-world drift: plugin catalogs moving ahead of pinned hosts, custom policies moving through snapshot restore, host-specific gateway behavior, and provider settings that should not leak from one inference runtime into another. The release keeps tightening the contract between onboarding, restore, messaging, and inference so the sandbox does the predictable thing even when the environment around it changes.

For users, the release improves Brave web search setup, custom policy restores, gateway health reporting, WSL and macOS behavior, Nemotron Ultra tool-less responses, provider switching, and Hermes WhatsApp pairing. For operators, it also moves more messaging and onboarding behavior into structured plans and smaller helper flows, which makes the setup pipeline easier to test and harder to accidentally regress.

The maintenance story is just as important here. v0.0.64 expands the Vitest live E2E migration, refreshes user docs and generated skills, makes selective E2E reporting more honest when jobs are cancelled, and keeps release confidence moving forward without turning internal cleanup into user-facing noise.

OpenClaw, Sandbox, and Onboarding

• #4955 pins the Brave web-search plugin to the OpenClaw host version and preserves the BRAVE_API_KEY placeholder through build doctor. This keeps Brave search from breaking when the upstream plugin catalog advances beyond NemoClaw's pinned OpenClaw runtime.
• #5104 captures and replays custom policy presets during snapshot restore. Custom policy --from-file and --from-dir presets now survive backup and restore instead of silently falling back to built-in policy state.
• #5116 ties the in-container gateway health marker to the actual gateway launch site. Docker-driver sandboxes avoid misleading unhealthy status when the gateway is not meant to run inside the container.
• #5204 and #5209 stabilize CLI dispatch and platform tests on WSL, macOS, and related host paths. These fixes reduce platform-specific churn around public command dispatch and Vitest coverage.
• #5269 updates the gateway rootfs compatibility guard after the Dockerfile patch flow moved into its own helper. The E2E guard now follows the code that actually passes the macOS VM-rootfs compatibility flag.
• #5125, #5128, #5130, #5134, #5136, #5140, #5154, #5158, #5164, #5165, #5171, #5173, #5175, and #5176 continue splitting onboarding into focused runtime, provider, build, registration, launch, and prompt-selection helpers. The visible behavior stays steady while the release reduces the size and risk of future onboarding changes.

Messaging and Runtime Activation

• #4949 moves messaging build setup into manifest hooks for OpenClaw and Hermes. Messaging package installs, static render outputs, and post-agent-install behavior now come from the staged plan instead of scattered build arguments and legacy scripts.
• #5123 shifts channel state toward plan-backed messaging data. The runtime can reason from the same compiled channel plan that drives build and rebuild behavior.
• #5135 and #5148 extract messaging preparation and build patch configuration from the larger onboarding flow. This makes channel setup easier to test independently and lowers the risk of future onboarding edits.
• #5210 recovers replaced OpenClaw scope-upgrade approvals. Pairing and pending-state recovery are more deterministic when approval state changes underneath a sandbox.
• #5268 refreshes stale messaging render plans during rebuild. Existing Hermes Discord and similar channel plans can regenerate missing render entries while preserving credential binding hashes.

Hermes and Inference

• #5085 adds a narrow Nemotron Ultra 550B system-prompt injection for requests with no system message and no execution-capable tools. The model can emit the concrete steps it was already planning instead of losing useful output in tool-less flows.
• #5241 forces the completions API for providers that do not expose /v1/responses, including nvidia-prod, nvidia-nim, and gemini-api. Provider switches no longer inherit an incompatible Responses API setting from a previous runtime.
• #5257 preinstalls Hermes WhatsApp bridge dependencies in the sandbox base image when the upstream lockfile is present. WhatsApp pairing no longer depends on writing node_modules into a read-only runtime path.

Docs, Skills, and Release Surface

• #5099 fixes the sandbox-hardening link and security layer count. The docs now match the current security framing.
• #5162 refreshes the managed-vLLM catalog skills. Agent-facing guidance stays aligned with the current inference catalog.
• #5196 adds a homepage starter prompt and improves the entry point for new users. The first interaction with the docs now gives users a clearer prompt for trying NemoClaw.
• #5244 refreshes v0.0.63 release docs and generated user skills. It carries forward recently shipped sandbox recovery, OpenClaw config restore, managed vLLM, Slack, and diagnostics guidance.
• #5247 fixes QA-reported command drift across deployment, sub-agent setup, command reference, and troubleshooting pages. The docs now point users toward the current installer, Docker-driver, and NemoClaw-managed recovery flows.

CI and Release Confidence

• #5049 adds failing-test-first coverage and recovery helpers for sandbox recovery work. The release keeps a regression guard around the behavior before broadening the implementation.
• #5147, #5185, #5186, #5187, #5189, #5218, #5219, #5223, #5224, #5226, #5228, #5231, #5232, #5233, #5234, #5236, and #5256 move more live and migration E2E coverage into Vitest. Double onboard, launch smoke, rebuild, sandbox operations, network policy, credential migration, inference routing, plugin EXDEV, TUI unreachable inference, token rotation, Hermes, dashboard binding, WhatsApp QR, VM driver, and docs validation paths now have a clearer test structure.
• #5170 adds an E2E suite review lens, and #5243 adds a Vitest job selector. Together they make the expanding live test suite easier to review and target.
• #5167 switches the PyPI policy probe to a real pythonhosted artifact. Policy validation now tests against a more realistic package-hosting path.
• #5203 and #5212 recover common egress OpenClaw retries and shard channel stop-start nightly coverage. These changes improve nightly signal around flaky sandbox and channel paths.
• #5246 buckets cancelled jobs in the Selective E2E Results comment. Cancelled runs now report as no-signal instead of accidentally appearing as green.
• #5183 and #5184 update GitHub Actions dependencies. The release keeps routine workflow dependencies current alongside the larger test migration.

Thank you

Thank you to external contributor @johnellison for the Brave web-search plugin and API-key preservation fix in #4955.

[convexchaos]

Any plans to bump OPENCLAW_VERSION to 2026.6.x in an upcoming release? The Telegram streaming fix (streamed text survives tool calls) and lower first-reply latency in 2026.6.6 would be especially valuable for DGX Spark users running large local models, where reply latency and mid-tool-call truncation are current pain points. Understand the Dockerfile target lags for stability validation, mostly asking about expected cadence.

[cv]

Upgrades to the OpenClaw, OpenShell and Hermes pins are planned soon, but we need to get a better way to run canary E2E builds first to guarantee they're not going to break (too many) things. This should land at some point in this week or the next!